What AI says about MA-RLHF

MA-RLHF is an open-source codebase implementing reinforcement learning from human feedback (RLHF) using macro actions to guide policy optimization for language models. The repository provides prerequisites and environments (Gemma variants, PyTorch, DeepSpeed, Transformers) and includes a DeepSpeed-Chat-based implementation to train and evaluate aligned AI agents, along with components like Constituent-Treelib. Developed by Ernie Research, the project was released around ICLR 2025 to enable reproducing the MA-RLHF approach for improved alignment in large language models.